Short answer
AI proof of delivery extraction reads signed delivery dockets, driver photos and electronic POD exports and turns each into a structured record: consignment number, delivery time, receiver, quantities and condition remarks. For New Zealand carriers and 3PLs it speeds invoicing and dispute handling, but it must never drop a damage remark, because those notes bear on claims under the Contract and Commercial Law Act 2017.
- Extract consignment number, time, receiver, quantity and remarks from every proof of delivery.
- Clean PODs release invoicing; mismatches and handwritten remarks go to a person.
- Delivery photos and GPS coordinates count as personal information under the Privacy Act 2020.
- A carrier’s responsibility ends when goods are tendered under the contract; the POD shows when.
What gets extracted, and from what
Proof of delivery extraction turns each delivery’s evidence into the same short record, whatever form it arrived in. Most New Zealand carriers handle at least three forms: paper consignment notes scanned at the depot, phone photos of signed dockets, and exports from an ePOD app.
| Field | Typical source | Why it matters |
|---|---|---|
| Consignment number | Barcode or printed number | Links the POD to the job, the invoice and any claim |
| Delivered date and time | App timestamp or handwritten time | Fixes when the carrier’s responsibility ended |
| Receiver name and signature | Signature box, or authority to leave flag | Shows who accepted the goods |
| Quantity received | Handwritten count against the manifest | Catches short deliveries before billing |
| Condition remarks | “1 ctn crushed”, “wet pallet”, driver notes | Starts the damage conversation |
| Photo check | Goods at a door or dock, address visible | Supports unattended deliveries |
How the extraction runs
- Inputs
- Depot scans of paper dockets, driver phone photos, ePOD exports (signature image, timestamp, GPS point, photos), and the manifest for each run.
- Systems
- A vision model for the images, your TMS or WMS for jobs and manifests, document storage for originals, and Xero or your billing system for invoice release.
- Steps
- Read the barcode or number; match to the job; extract fields into a fixed schema; compare quantities with the manifest; detect any writing outside the signature box; check the photo shows goods; score confidence per field.
- Human review
- An administrator clears every POD with a remark, a quantity mismatch, a missing signature without authority to leave, or a low-confidence field. A supervisor audits a weekly sample of auto-accepted PODs.
- Outputs
- A structured POD record on the job, the original image linked, invoice release for clean deliveries, and a remarks queue feeding shipment exception summaries.
- Measures
- Share of PODs auto-accepted, error rate in the weekly sample, hours of keying removed, days from delivery to invoice, and disputes where the POD could not be found.
- Risks
- Remarks lost in extraction, misread quantities, a docket matched to the wrong job, and delivery photos stored with no retention limit.
Illustrative example: a regional carrier’s week
Illustrative example: a Christchurch regional carrier making 2,400 deliveries a week. Assumptions: 700 arrive as paper dockets keyed by hand at 1.5 minutes each; 85% of paper dockets pass every check once tuned (a figure to prove in a pilot, not a promise); 3% of ePOD deliveries carry a free-text remark. All numbers are synthetic.
| Work item | Current hours a week | With extraction (assumed) |
|---|---|---|
| Keying 700 paper dockets | 17.5 | None |
| Reviewing 105 flagged dockets at 1.5 minutes | None | 2.6 |
| Reading 51 ePOD remarks at 1 minute | Rarely done | 0.9 |
| Auditing 20 auto-accepted PODs | None | 0.3 |
| Total | 17.5 | 3.8 |
Model cost depends on the model you need for handwriting. Assume 3,000 input and 300 output tokens per docket image. At Claude Haiku 5.5 list prices (US$0.10 and US$0.50 per million tokens) that is about US$0.30 a week; at Claude Sonnet 5.5 (US$2 and US$10) about US$6.30, as listed on 19 October 2026. Choose by accuracy on your dockets, not by price.
The New Zealand rules behind a POD
When responsibility ends. Under the Contract and Commercial Law Act 2017, a contracting carrier’s responsibility ends when goods are tendered to the consignee in accordance with the contract, or for collections when collected or five days after notice they are ready (section 258). The extracted timestamp is the evidence of that moment, so it has to be right.
Damage remarks. Written notice of damage or partial loss is due within 30 days after responsibility ends (section 274), but notice is not required if the carrier is or ought to be aware of the damage (section 276). A remark written on your own docket bears directly on that question. Extraction must surface it, never summarise it away.
Photos and locations. The Privacy Commissioner’s AI and privacy guidance says personal information includes photographs of a person and metadata such as map coordinates. Door photos and GPS points qualify. Principle 9 says not to keep personal information longer than required; the Act’s 12-month limit on proceedings (sections 278 and 279) is a sensible floor for delivery evidence.
What stays with people
- Accepting a POD for a high-value consignment where the signature is missing or the name unreadable.
- Deciding whether a remark becomes a claim, and telling the customer. That conversation follows our customer delivery enquiry handling approval rules.
- Releasing any invoice the system held back.
- Changing the extraction schema or the auto-accept thresholds.
Failure modes we design around
- Remarks flattened. “Rec’d 9 of 10, 1 ctn crushed” becomes “delivered”. Control: a separate check for any writing outside the signature box; a non-empty remark always goes to review.
- Misread digits. A handwritten 7 read as 1. Control: compare every quantity with the manifest and never auto-accept a mismatch.
- Wrong job. A driver photographs yesterday’s docket. Control: match number, date and address together.
- Time zone drift. ePOD exports often use UTC, which can move an evening delivery to the next day in New Zealand time. Control: convert and store local time with the zone.
- Camera roll copies. Photos stay on drivers’ personal phones. Control: capture inside the app and keep originals in managed storage.
Choosing between OCR and a vision model is its own decision; see OCR versus vision models. The weekly sample audit works well with our AI quality review checklist. For build options see AI document processing, for billing AI and Xero, and for the sector view AI for logistics in New Zealand.
Sources
- Contract and Commercial Law Act 2017 (Part 5, subpart 1: carriage of goods), New Zealand Legislation, 2017
- Artificial intelligence and the Information Privacy Principles, Office of the Privacy Commissioner, 2023
- Principle 9: Retention of personal information, Office of the Privacy Commissioner, accessed 19 October 2026
- Claude API pricing, Anthropic, accessed 19 October 2026
Written by the Levy Street team in Wellington. Last updated 19 October 2026.